Where AI learns to do real work

Labelbox builds the environments frontier labs train in and the platform enterprises run their agents on.

Recursion· Managed AgentsLabelbox's platform for running agents on enterprise work

On Recursion, a team of eight agents reads out a finished 48-run training sweep.

Across the tracker, cluster, evals, GitHub, Notion, Slack and Docs, each agent on the model it is best at. A grader agent scores the outcome.

sweep-2291 · curriculum-v3 · 48 runs · 16 configs × 3 seeds · 1.3B

Readout: does curriculum-v3 beat the baseline on reasoning evals at 1.3B?

Check run health, verify the evals, compare to baseline across seeds, and promote the winning configs to the 7B run.

running

8 agents · 4 models

outcome 0.91 · 2 configs promoted · 7B launched

Tracker0
Cluster0
Evals0
GitHub0
Notion0
Slack0
Docs0

coordinator · Tracker: 48 runs, 16 configs × 3 seeds · plan: 7 tasks across 4 specialists

For enterprises

Agents that do the work, and get better at it.

Recursion Managed Agents are a new way to use AI in the business. Set a goal; Recursion plans it, spawns as many agents as the problem needs, and works it in parallel across your systems, autonomously. Every run makes the next one better.

Recursion

The managed agent platform for the hardest work in your company. One goal in; a coordinator, as many specialists as it takes, and one graded outcome out.

Set the goal, not the steps

Hand Recursion a goal: retrain a model on this week's data, underwrite a loan, close the quarter. A coordinator plans it and works across your systems on any model.

Hard problems get more agents

When the problem is big, Recursion spawns specialists to attack it in parallel: persistent, long-horizon, asynchronous agents running as one fleet. A grader scores every outcome.

The fleet remembers

Graded sessions become memories, notes and skills that later agents recollect before they act. The fleet gets better every week, without retraining.

Fine-tune on what the fleet learns

Turn the fleet's graded runs into fine-tuning and RL on Labelbox Horizon, the frontier-class RL environments trusted by over 90% of leading US AI labs. Fine-tuned models deliver a 5–10× lower cost per task.

For frontier AI

The data, environments, and evaluation infrastructure the world's frontier AI labs build on.

Horizon

RL environments and preference signals tuned for post training and evals, spanning reasoning, tool use, and computer use across autonomous AI research, scientific knowledge work, agent coding, and cybersecurity.

Terra

Full stack data products for robotics foundation models. Video, trajectories, and rich multimodal annotations, collected with purpose-built hardware and AI-powered diversity engines, spanning pre-training, post-training, and evaluation.

Alignerr

Real-world grounding signal from the smartest people on earth — the kind frontier models can't generate themselves. Alignerr brings together 2.6M+ knowledge experts, channeling human endeavor into the reward signals frontier AI needs to be useful.

Hundreds of AI teams build with Labelbox

How Meta built GIM with Labelbox data to evaluate frontier AI reasoning

Problem

Meta needed a benchmark that remained discriminative as existing LLM evaluations saturated. The team wanted tasks grounded in practical reasoning rather than obscure knowledge or synthetic puzzles, with enough rubric detail to capture partial credit and enough quality control to support a public-private contamination diagnostic.

Solution

Labelbox produced the data foundation for GIM: 820 expert-authored problems across seven cognitive categories, including 229 multimodal items and 528 rubric-graded prompts. The work included original prompt creation, structured scoring criteria, review, annotation, and quality assurance, enabling Meta to calibrate a 2PL IRT model over more than 200,000 prompt-response pairs.

Result

Meta released GIM-615, calibrated item parameters, and an evaluation framework that benchmarked 22 models across 47 reporting configurations. The paper found GIM remains far from saturated, with roughly 20% of items above frontier ability, giving researchers a durable way to compare model capability, thinking budgets, and future systems.

Labelbox Research

Latest work from Labelbox Research

Labelbox's world-class applied research team pioneers frontier AI data generation and evaluation methods. Through scientific precision and co-innovation, we help customers achieve real-time AGI breakthroughs.

EchoChain

An audio benchmark for reasoning under pressure in full-duplex dialogue

Implicit intelligence

Evaluating agents on what users don’t say

Intent laundering

The AI safety illusion: why current safety datasets fool us on model safety